Fix with AI: One Click When a Publish Fails

When publishing fails, Mythex now says what happened and offers Fix with AI: the agent reads the build output or your app's logs and fixes the project.

Mythex Team · 2026-09-29 · 4 min read

When a publish fails on Mythex, the dialog now tells you what actually happened and gives you two buttons on every failure: Fix with AI and Try again. Fix with AI hands the agent the failure and what your app printed, so it can fix the project and publish again — and if nothing in your app was wrong, it simply republishes. We also stopped blaming "our side" for problems that were really in the app, and started catching apps that break a few minutes after going live.

What changed

Fix with AI on every failure

Before, some failures showed Fix with AI and others only showed Try again. The split was based on whose fault the failure looked like: if it seemed to be on our side, we didn't want the agent chasing code that was fine, on your credits.

That left you to work out which kind of failure you were looking at. Now every failure offers both buttons. The agent gets the failure's type and the output behind it, so it fixes the app when something in the app is wrong, and republishes when nothing is. You never have to tell which it was.

The exceptions are the ones where the answer isn't code. A taken link name suggests other names. Running out of credits offers a way to get more.

Messages that say what happened

Several failures used to end up as "a problem on our side, publish again" — which was untrue, and publishing again never helped. Now they have their own messages:

"Your app didn't start after publishing." The app was published but never began accepting connections. Most often it listens on a different port, or only on localhost. A published app has to listen on 0.0.0.0 and the port in $PORT. Instead of waiting a fixed time and hoping, we now check that the app actually accepts connections on its port.
"This file is too big" / too many files. Static sites have a per-file size limit (about 25–30 MB, depending on the kind of site) and a cap of 20,000 files. We now measure the folder before uploading anything, the same way for every kind of file, and the message lists every file over the limit.
"Your app stopped right after it started." Usually a missing setting such as an environment variable or secret, a crash on the first request, or an error in the start command.

Catching apps that break after going live

A publish can succeed and the app still fail a few minutes later: it keeps crashing, or it runs but stops accepting connections. Mythex checks published apps regularly. If an app crashes repeatedly within a few minutes, or is running but taking no connections, the publish is marked failed, you get an email, and the project shows the message with Fix with AI.

A single crash that the app recovered from on its own doesn't count as a failed publish. If it was because the app ran out of memory, you get a separate email about that.

When we mark a live app as failed, we also save what the app printed at that moment. You never see that raw output, but it's what Fix with AI gives the agent, so it starts from the real error instead of guessing.

How to use it

Click Publish as usual.
If it fails, read the message in the dialog. It's written to tell you what happened, not just that something did.
Click Fix with AI. The agent reads the build output or the app's logs, fixes the project, and you can publish again.
If the message says nothing is wrong with your app — "Publishing is busy for a moment" or "Publish was interrupted" — Try again is usually enough.
If you got an email that a live app keeps crashing, open the project: the same message and Fix with AI are waiting there.

Keep the same link name when you publish again, unless the message says the name is the problem.

Why we built it

A failed publish is often the first real wall someone hits. The app works in the preview, they press Publish, and it doesn't come up. For someone who isn't a developer, "check your logs" is not advice they can act on. Fix with AI turns that moment into one click, and honest messages mean that when you do read them, they point at the right thing.

Limits

Fix with AI uses the agent, so it uses credits like any other message.
It can fix what's in your project: code, configuration, the start command. It can't fix a taken link name, an unverified email or an empty credit balance.
Some problems need a decision from you, like whether to move a large file into Storage or make it smaller.

Every message and what to do about it is in the docs: Publish troubleshooting. For the broader skill of fixing things when they go wrong, see how to debug an AI-built app.